loopx

LoopX is a local-first semantic control plane and research workbench for long-running agent teams. Its premise is that a single session can finish a task, but long work breaks down when objectives change, owner decisions appear, evidence goes stale and agents hand work to peers. LoopX keeps the goal, accepted requirements, decisions, ownership and evidence durable across sessions, restarts and runtimes; your agent host supplies the model and tools.

The README describes it as an agent-native Kanban for long-running work: cards carry identity, authority, evidence and continuation, and moves are validated operators such as claim, gate, monitor and writeback. Registered agents are peers coordinated by claims, leases and typed continuation rather than a fixed leader. It is pitched at multi-day engineering, research and benchmark objectives, PR loops that must preserve scope and review state, and heartbeat or monitor work. It is explicitly not an autonomous production controller: dangerous permissions and publishing stay with the human.

Install with pip install loopx and loopx workflow-skills --install, then call $loopx in Codex or /loopx in Claude Code; Python 3.11+ and Node.js 22.22.3+ are required. The project published a 46-task study on Long-Horizon Terminal-Bench in which LoopX Heartbeat reached a higher mean reward than plain Codex and native Codex Goal with the same model, and the README spells out the study's limits. Apache-2.0 licensed.